This is the stage where code actually changes, and it happens inside a dedicated Git worktree and branch so the work is isolated from the trunk until it is ready. A worker agent drives it test-first: one smallest failing check, then the minimal implementation that makes it pass, then the next. Before writing the happy-path implementation the worker walks an unhappy-path checklist, empty and zero and missing inputs, reload and persistence, concurrency, malformed input, and tests each failure that could actually violate acceptance.

open ⤢
Unhappy paths first, then red to green, one smallest test at a time. The diff budget is a circuit breaker and the real fixture is the last check before review.

The failing test built first is the anchor. Its job is to prove the check is real by watching it go red, so that green afterwards means something rather than meaning the test was written to match whatever the code already did. A fix that lands one commit after its feature is the tell that the tests were written to the demo instead of to the requirement. The diff budget acts as a circuit breaker, not a target, and the finished change is run against the real fixture named in the work unit, not a synthetic stand-in, because passing invented inputs while the real one breaks is exactly the failure this stage exists to catch.